A smart cleaning system employing robotic arms for cleaning and drying, and its working method.

Abstract

This invention relates to the field of control system data processing technology, and in particular to an intelligent cleaning system and its working method that employs a robotic arm for cleaning and drying. The system includes a visual data acquisition module that acquires raw image data; an environmental adaptation processing module that outputs an optimized image matrix; a feature recognition module that extracts part contour and position features from the optimized image matrix and outputs coordinate and posture information; a control decision module that fuses the feature information with preset cleaning parameters to generate robotic arm motion commands; and a robotic arm execution module that drives the spray gun to complete the cleaning action and feeds back the position data acquired by the joint encoder to the feature recognition module to form a closed-loop control. This invention compensates for low-light conditions and suppresses stain noise through environmental adaptation processing, enhances robustness through feature recognition combined with a confidence mechanism, and corrects recognition deviations through feedback loops, ultimately achieving accurate detection of part positions and complete cleaning, thus solving the problems of missed cleaning and uneven cleaning.

Description

Technical Field

[0001] This invention relates to the field of control system data processing technology, and in particular to an intelligent cleaning system and its working method that uses a robotic arm for cleaning and drying. Background Technology

[0002] The intelligent cleaning system integrates robotic arms and air-drying technology to achieve a fully automated cleaning process. The robotic arm executes precise movements through programming control to clean the surface of objects, such as spraying cleaning agents and brushing, to remove stains. Subsequently, the air-drying system uses high-speed airflow circulation to evaporate residual moisture, completing the drying stage, thereby improving the continuity and reliability of the cleaning process.

[0003] Existing intelligent cleaning systems suffer from the following technical pain points in terms of control: Specifically, vision sensors may malfunction in low-light or oily environments, leading to inaccurate part position detection and resulting in missed or uneven cleaning. This is because environmental interference, such as oil buildup or insufficient lighting, weakens the optical acquisition capabilities of the sensors, affecting the accuracy of image processing algorithms. For example, in intelligent cleaning workshops, high-pressure oil pump parts often have thick oil stains or are located in shadowed areas. Due to lens contamination or light scattering, vision sensors cannot accurately identify the edge contours of the parts. The control system adjusts the incident angle and trajectory of the robotic arm spray gun based on incorrect position data, causing the spray jet to deviate from the target area. Some parts are not covered while other areas are repeatedly cleaned, ultimately reducing the quality of the operation and increasing resource consumption. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent cleaning system and its working method that employs a robotic arm for cleaning and drying. This invention solves the technical problem of inaccurate part position detection caused by recognition errors of vision sensors in low light or oily environments, resulting in missed cleaning or uneven cleaning.

[0005] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows:

[0006] In a first aspect, the present invention provides an intelligent cleaning system that employs a robotic arm for cleaning and drying, comprising:

[0007] The visual data acquisition module is configured to scan the parts to be cleaned using a visual sensor installed at the end of the robotic arm, thereby acquiring raw image data.

[0008] The environmental adaptation processing module is connected to the visual data acquisition module and is configured to receive raw image data, perform light intensity analysis and stain area detection on the raw image data to obtain analysis results and stain area detection results, use the analysis results and stain area detection results to drive image enhancement and filtering processing, and output an optimized image matrix.

[0009] The feature recognition module is connected to the environment adaptation processing module and is configured to receive the optimized image matrix, extract the contour and position features of the part from the optimized image matrix, and output feature information including coordinates and attitude angles.

[0010] The control decision module is connected to the feature recognition module and is configured to receive feature information, call the system's preset cleaning parameters based on the feature information, fuse the feature information with the preset cleaning parameters, perform trajectory planning based on the fused data to calculate the robotic arm's motion trajectory, perform error compensation calculation to correct motion errors, and generate robotic arm motion commands.

[0011] The robotic arm execution module is connected to the control decision module and is configured to execute motion commands to drive the spray gun to complete the cleaning action. During the driving process, the joint encoder collects position data in real time and feeds the position data back to the feature recognition module.

[0012] Furthermore, in the intelligent cleaning system of the present invention, which employs robotic arms for cleaning and drying, the visual data acquisition module is also configured as follows:

[0013] Initiate the self-test program of the vision sensor to check the lens cleanliness and light source status;

[0014] After the self-test is completed, the vision sensor is controlled to scan the parts to be cleaned. During the scanning process, the oil and dirt on the lens surface are removed by the periodically activated automatic cleaning device.

[0015] During the scanning process, the ambient light intensity is dynamically analyzed, and the exposure parameters are automatically adjusted when the brightness is below the threshold to complete image acquisition;

[0016] The acquired raw image data is transmitted to the environmental adaptation processing module via the industrial Ethernet protocol.

[0017] Furthermore, the intelligent cleaning system of the present invention, which employs robotic arms for cleaning and drying, also includes an environmental adaptation module configured as follows:

[0018] The received raw image data is parsed, and the average brightness distribution of the raw image data is calculated.

[0019] Based on the average brightness distribution, when the average brightness of the original image data is lower than the threshold, the contrast enhancement algorithm is triggered to adjust the gamma value of the original image data.

[0020] An adaptive median filter is applied to scan the pixels of the original image data, identify noisy pixels in the image and replace them with the mean of neighboring pixels to obtain the filtered image.

[0021] The filtered image is subjected to color balance correction to obtain a color balance corrected image. The color balance corrected image is then converted into a grayscale matrix format and output to the feature recognition module.

[0022] Furthermore, in the intelligent cleaning system of the present invention, which employs robotic arms for cleaning and drying, the feature recognition module is further configured as follows:

[0023] Receive the optimized image matrix output by the environment adaptation processing module, perform background segmentation on the optimized image matrix, and distinguish between the part area and the background;

[0024] The segmented image is input into a pre-trained convolutional neural network model, and edge and contour features are extracted through convolutional and pooling layers;

[0025] The extracted features are classified using a fully connected layer to obtain the center coordinates and pose angles of the part; a confidence score is calculated based on the classification data of the center coordinates and pose angles.

[0026] If the confidence score is lower than the preset value, a re-collection command will be triggered;

[0027] If the confidence score is higher than the preset value, the center coordinates, attitude angle, and confidence score are encapsulated into a structured data format and transmitted to the control decision module.

[0028] Furthermore, in the intelligent cleaning system of the present invention, which employs robotic arms for cleaning and drying, the control decision module is further configured as follows:

[0029] Parse the received feature information to obtain the part coordinates and orientation;

[0030] Combining the part coordinates and orientation with the system's preset spray gun incident angle and cleaning time parameters, the inverse kinematics algorithm is used to calculate the motion trajectory of the robotic arm joints;

[0031] After calculating the motion trajectory, the part coordinates and posture are compared with historical recognition data. If the deviation exceeds the preset tolerance value, the coverage of the motion trajectory is adjusted to obtain the adjusted motion trajectory.

[0032] Based on the adjusted motion trajectory, the speed curve of the robotic arm's motion is optimized to generate smooth acceleration commands;

[0033] The motion trajectory and smooth acceleration commands are encapsulated into control signal packets and sent to the robotic arm execution module via the fieldbus protocol.

[0034] Furthermore, in the intelligent cleaning system of the present invention that employs a robotic arm for cleaning and drying, the robotic arm execution module is further configured as follows:

[0035] Decode the received control signal packets to obtain the target's position, velocity, and acceleration parameters;

[0036] Based on the target position, velocity, and acceleration parameters, drive the servo motor and reducer to control the movement of the robotic arm joints;

[0037] During the movement of the robotic arm joints, the spray gun pressure is adjusted to the preset range of the system, the flow sensor data is monitored, and the joint encoder position data of the robotic arm is collected in real time.

[0038] The joint encoder position data is sent as feedback data to the feature recognition module;

[0039] After receiving the feedback data, the feature recognition module compares the feedback data with the target position of the robotic arm. If the deviation between the feedback data and the target position of the robotic arm exceeds the system's preset tolerance value, it sends a fine-tuning command to the robotic arm execution module.

[0040] The robotic arm execution module receives a fine-tuning instruction, triggering a position fine-tuning mechanism to re-execute the cleaning action.

[0041] Furthermore, the intelligent cleaning system of the present invention, which employs robotic arms for cleaning and drying, also includes:

[0042] The optimized image matrix output by the environment adaptation processing module is used as the input to the convolutional neural network model in the feature recognition module.

[0043] Convolutional neural network models extract features from the optimized image matrix through convolutional layers and pooling layers, and output feature information.

[0044] Furthermore, the intelligent cleaning system of the present invention, which employs robotic arms for cleaning and drying, includes a confidence score in the feature information output by the feature recognition module.

[0045] The control decision module receives feature information and executes the following: if the confidence score is higher than the preset value, it generates robotic arm movement instructions based on the feature information.

[0046] If the confidence score is lower than the preset value, a re-acquisition command is triggered or historical recognition data is called to perform error compensation calculation, generating robotic arm movement commands with expanded coverage.

[0047] Furthermore, the intelligent cleaning system of the present invention, which employs robotic arms for cleaning and drying, also includes:

[0048] The robotic arm execution module feeds back the position data collected by the joint encoder to the feature recognition module;

[0049] The feature recognition module receives feedback position data, calculates the Euclidean distance between the feedback position data and the part coordinates obtained in the current recognition cycle as the position deviation, calculates the difference between the feedback position data and the part attitude angle as the attitude deviation, and merges the position deviation and attitude deviation to generate pose deviation data.

[0050] The feature recognition module adjusts the parameters of the convolutional neural network model in the feature recognition module based on the pose deviation data.

[0051] Secondly, the present invention provides an intelligent cleaning method employing a robotic arm for cleaning and drying, applicable to the intelligent cleaning system employing a robotic arm for cleaning and drying as described above, comprising:

[0052] Step 1: Scan the part to be cleaned to obtain raw image data; perform a self-check to inspect lens cleanliness and light source status; remove oil stains from the lens surface during scanning; dynamically analyze ambient light intensity and adjust exposure parameters when the brightness is below the threshold; transmit raw image data after image acquisition is complete.

[0053] Step 2: Receive the original image data, perform light intensity analysis and stain area detection, calculate the average brightness distribution of the original image data, trigger the contrast enhancement algorithm to adjust the gamma value when the average brightness of the original image data is lower than the threshold, apply adaptive median filtering to identify and replace noisy pixels, and output the optimized image matrix after performing color balance correction.

[0054] Step 3: Receive the optimized image matrix for background segmentation, extract contour features through convolutional and pooling layers, classify and output part coordinates and pose through a fully connected layer, and calculate the confidence score; if the confidence score is lower than the set value, trigger re-acquisition; if it is higher than the set value, encapsulate the coordinates, pose, and confidence score into structured data.

[0055] Step 4: Receive the encapsulated structured data as feature information, parse the part coordinates and attitude, combine the system's preset spray gun incident angle and cleaning time parameters, use the inverse kinematics algorithm to calculate the motion trajectory, compare historical identification data and adjust the trajectory coverage when the deviation exceeds the limit, optimize the speed curve to generate smooth acceleration commands, encapsulate the motion commands and send them.

[0056] Step 5: Decode the control signal packet to obtain motion parameters, drive the movement of the robotic arm joints, adjust the spray gun pressure and monitor the flow data, and collect the joint position data of the robotic arm in real time as feedback data; compare the feedback data with the target position of the robotic arm, and send a fine-tuning command to trigger a re-cleaning action when the deviation between the feedback data and the target position of the robotic arm exceeds the limit.

[0057] Step 6: Based on the feedback data, calculate the Euclidean distance between the feedback data and the part coordinates as the position deviation, calculate the difference between the feedback data and the attitude angle as the attitude deviation, merge to generate pose deviation data, and adjust the parameters of the convolutional neural network model.

[0058] Beneficial effects of this invention:

[0059] The beneficial effects of this invention lie in its effective solution to the technical problem of inaccurate part position detection caused by recognition errors of vision sensors in low-light or oily environments through multi-module collaborative data processing and closed-loop control mechanisms. The vision data acquisition module initiates a self-test program to check the cleanliness of the lens and the state of the light source. During the scanning process, an automatic cleaning device removes oil stains from the lens and dynamically adjusts exposure parameters to acquire high-quality raw image data. The environmental adaptation processing module performs light intensity analysis and stain area detection, drives image enhancement and filtering, improves the visibility of details in dark areas through contrast enhancement algorithms, suppresses noise pixels through adaptive median filtering technology, and compensates for color cast through color balance correction, outputting an optimized image matrix. The feature recognition module uses a convolutional neural network model to extract the contour and position features of the part, outputting coordinates and orientation angles. The system collects feature information and confidence scores. If the confidence score is lower than a set value, a re-acquisition command is triggered to avoid misidentification. The control decision module integrates feature information with preset cleaning parameters, generates robotic arm motion commands through trajectory planning and error compensation calculations, and dynamically adjusts the motion trajectory coverage by comparing the current feature information with historical recognition data. The robotic arm execution module drives the spray gun to complete the cleaning action and feeds back the position data collected by the joint encoder to the feature recognition module, forming a closed-loop control to correct recognition deviations in real time. The system compensates for low lighting conditions through image enhancement, suppresses stain noise through filtering, filters low-quality recognition results through a confidence mechanism, expands the motion range through historical data compensation, and corrects execution errors through feedback closed-loop correction, ultimately improving the accuracy of part position detection, reducing missed cleaning and uneven cleaning, and achieving efficient, reliable, and intelligent cleaning. Attached Figure Description

[0060] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0061] Figure 1 The flowchart below illustrates an intelligent cleaning system and its working method that employs a robotic arm for cleaning and drying. Detailed Implementation

[0062] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention provided by various embodiments will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0063] In a first aspect, the present invention provides an intelligent cleaning system that employs a robotic arm for cleaning and drying, comprising:

[0064] The visual data acquisition module is configured to scan the parts to be cleaned using a visual sensor installed at the end of the robotic arm, thereby acquiring raw image data.

[0065] The environmental adaptation processing module is connected to the visual data acquisition module and is configured to receive raw image data, perform light intensity analysis and stain area detection on the raw image data to obtain analysis results and stain area detection results, use the analysis results and stain area detection results to drive image enhancement and filtering processing, and output an optimized image matrix.

[0066] The feature recognition module is connected to the environment adaptation processing module and is configured to receive the optimized image matrix, extract the contour and position features of the part from the optimized image matrix, and output feature information including coordinates and attitude angles.

[0067] The control decision module is connected to the feature recognition module and is configured to receive feature information, call the system's preset cleaning parameters based on the feature information, fuse the feature information with the preset cleaning parameters, perform trajectory planning based on the fused data to calculate the robotic arm's motion trajectory, perform error compensation calculation to correct motion errors, and generate robotic arm motion commands.

[0068] The robotic arm execution module is connected to the control decision module and is configured to execute motion commands to drive the spray gun to complete the cleaning action. During the driving process, the joint encoder collects position data in real time and feeds the position data back to the feature recognition module.

[0069] This invention relates to an intelligent cleaning system employing a robotic arm for cleaning and drying, achieving an automated cleaning process through multi-module collaboration. A visual data acquisition module scans the parts to be cleaned using a vision sensor mounted at the end of the robotic arm, thereby acquiring raw image data. Specifically, the vision sensor initiates a self-test program to check lens cleanliness and light source status. After the self-test is completed, it controls the vision sensor to execute the scanning trajectory. During the scanning process, an automatic cleaning device is periodically activated to remove oil and dirt from the lens surface, preventing image quality degradation. Simultaneously, the ambient light intensity is dynamically analyzed, and when the brightness falls below a preset threshold, exposure parameters such as exposure time and ISO value are automatically adjusted to optimize acquisition conditions. The acquired raw image data is transmitted to the environmental adaptation processing module via an industrial Ethernet protocol, achieving efficient and reliable data exchange. This step provides high-quality input data for subsequent processing, connecting to the environmental adaptation processing module.

[0070] After receiving the raw image data, the environmental adaptation processing module performs illumination intensity analysis and stain area detection. Specifically, it parses the raw image data and calculates the average brightness distribution of the image. When the average brightness is below a threshold, it triggers a contrast enhancement algorithm to adjust the image gamma value, improving the visibility of details in dark areas. It then applies adaptive median filtering to scan image pixels, identifies noisy pixels, and replaces them with the average of neighboring pixels to suppress stain interference. Finally, it performs color balance correction to compensate for color cast caused by oil stains, obtaining a filtered image. The processed image is then converted to a grayscale matrix format and output to the feature recognition module. This invention compensates for environmental interference through image enhancement and filtering, based on the output of the visual data acquisition module, and provides optimized input to the feature recognition module.

[0071] The feature recognition module receives the optimized image matrix and performs background segmentation to distinguish the part area from the background. The segmented image is then input into a pre-trained convolutional neural network model, which extracts edge and contour features through convolutional and pooling layers. A fully connected layer classifies the extracted features, outputting the center coordinates and pose angle of the part. A confidence score is calculated based on the classification data. If the confidence score is lower than a set value, a re-acquisition command is triggered; if it is higher, the coordinates, pose angle, and confidence score are encapsulated in a structured data format and transmitted to the control decision module. This process achieves accurate feature extraction through the convolutional neural network model, relies on the output of the environment adaptation processing module, and improves recognition reliability through a confidence mechanism.

[0072] After receiving the feature information, the control decision module analyzes the part's coordinates and posture. Combining this with the system's preset spray gun incident angle and cleaning time parameters, it uses an inverse kinematics algorithm to calculate the motion trajectory of the robotic arm joints. After calculating the trajectory, it compares the current feature information with historical recognition data. If the deviation exceeds a tolerance value, it dynamically adjusts the coverage area of ​​the trajectory. It optimizes the robotic arm's motion speed curve to generate smooth acceleration commands. The motion trajectory and smooth acceleration commands are encapsulated into a control signal packet and sent to the robotic arm execution module via a fieldbus protocol. This invention achieves accurate generation of motion commands through trajectory planning and error compensation, based on the output of the feature recognition module, and connects to the robotic arm execution module.

[0073] The robotic arm execution module decodes control signal packets to obtain target position, velocity, and acceleration parameters, driving servo motors and reducers to control the movement of the robotic arm joints. During movement, it adjusts the spray gun's spray pressure to a set range and monitors flow sensor data. Real-time acquisition of joint encoder position data is sent as feedback data to the feature recognition module. The feature recognition module compares the feedback data with the target position, sending a fine-tuning command when the deviation exceeds a tolerance value. The robotic arm execution module triggers a position fine-tuning mechanism to re-execute the cleaning action. These steps form a closed-loop control system, based on commands from the control decision module and using feedback to correct recognition deviations in real time, thus improving cleaning accuracy.

[0074] The system achieves closed-loop optimization from image acquisition to execution feedback through data flow. Each module step is connected sequentially. Environmental adaptation processing compensates for low light conditions, feature recognition combined with confidence mechanism filters data, control decision fusion parameters generate instructions, and robotic arm execution and feedback ensure accurate action, ultimately solving the problems of missed washing and uneven washing.

[0075] Specifically, the intelligent cleaning system of the present invention, which employs robotic arms for cleaning and drying, further includes a visual data acquisition module configured as follows:

[0076] Initiate the self-test program of the vision sensor to check the lens cleanliness and light source status;

[0077] After the self-test is completed, the vision sensor is controlled to scan the parts to be cleaned. During the scanning process, the oil and dirt on the lens surface are removed by the periodically activated automatic cleaning device.

[0078] During the scanning process, the ambient light intensity is dynamically analyzed, and the exposure parameters are automatically adjusted when the brightness is below the threshold to complete image acquisition;

[0079] The acquired raw image data is transmitted to the environmental adaptation processing module via the industrial Ethernet protocol.

[0080] The visual data acquisition module initiates a self-test procedure for the visual sensor, checking lens cleanliness and light source status. The self-test procedure includes capturing a built-in test image, analyzing image contrast and sharpness to assess the degree of lens contamination, and simultaneously using a photometer to check if the light source intensity is within its normal operating range. This self-test ensures the visual sensor is in optimal working condition, providing a reliable foundation for subsequent image acquisition and seamlessly connecting to the scanning process.

[0081] After the self-test is completed, the vision data acquisition module controls the vision sensor to scan the part to be cleaned. During the scanning process, an automatic cleaning device is periodically activated, for example, by using pneumatic spray cleaning agent to remove oil stains from the lens surface, preventing oil stains from adhering and affecting image quality. This process maintains lens transparency through real-time cleaning, ensuring continuous image acquisition during the scanning process, and triggering the scanning action based on the self-test results.

[0082] During the scanning process, the visual data acquisition module dynamically analyzes the ambient light intensity, monitoring the brightness value in real time via an ambient light sensor. When the brightness falls below a preset threshold, exposure parameters, including exposure time and ISO gain, are automatically adjusted to optimize image acquisition conditions. After image acquisition is complete, high-quality raw image data is obtained. This process improves data quality by using adaptive exposure to compensate for low-light interference, combined with cleaning measures.

[0083] The visual data acquisition module transmits the acquired raw image data to the environmental adaptation processing module via the industrial Ethernet protocol. Data transmission uses the industrial Ethernet protocol to encapsulate image data packets, achieving high-speed, low-latency, and reliable communication. This step completes the initial flow of image data, providing input to the environmental adaptation processing module and forming the starting point of the system data chain.

[0084] Specifically, the intelligent cleaning system of the present invention, which employs robotic arms for cleaning and air drying, further includes an environmental adaptation module configured as follows:

[0085] The received raw image data is parsed, and the average brightness distribution of the raw image data is calculated.

[0086] Based on the average brightness distribution, when the average brightness of the original image data is lower than the threshold, the contrast enhancement algorithm is triggered to adjust the gamma value of the original image data.

[0087] An adaptive median filter is applied to scan the pixels of the original image data, identify noisy pixels in the image and replace them with the mean of neighboring pixels to obtain the filtered image.

[0088] The filtered image is subjected to color balance correction to obtain a color balance corrected image. The color balance corrected image is then converted into a grayscale matrix format and output to the feature recognition module.

[0089] The environmental adaptation processing module parses the received raw image data and calculates the average brightness distribution of the raw image data. Specifically, the parsing operation includes reading the pixel values ​​of the raw image data, calculating the average brightness of all pixels, and generating a brightness distribution histogram to characterize the overall illumination conditions of the image. This step provides basic data for subsequent processing by quantitatively analyzing the image brightness characteristics, based on the raw image data transmitted by the visual data acquisition module, and provides input for brightness judgment.

[0090] Based on the average brightness distribution, when the average brightness of the original image data is lower than a preset threshold, the environment adaptation processing module triggers a contrast enhancement algorithm to adjust the gamma value of the original image data. The contrast enhancement algorithm improves the visibility of details in dark areas by mapping pixel brightness values ​​through nonlinear transformation; gamma value adjustment corrects the image grayscale curve to compensate for contrast loss caused by low light. This step adaptively enhances image quality, relying on the analysis results of the average brightness distribution, and prepares an optimized image for filtering processing.

[0091] The environmental adaptation processing module applies adaptive median filtering to scan the pixels of the original image data, identifying noisy pixels and replacing them with the average of neighboring pixels. The adaptive median filtering dynamically adjusts the filtering window size, traverses the image pixels to detect outliers, and replaces noisy pixels with the weighted average of surrounding pixels, suppressing dirt or random noise interference. This step improves image smoothness through local filtering, based on the contrast-enhanced image data, and provides clean input for color correction.

[0092] Color balance correction is performed on the filtered image to obtain the color-balanced image. Color balance correction analyzes the distribution deviation of the image's color channels and adjusts the gain ratios of the red, green, and blue channels to compensate for color cast caused by oil stains. Finally, the color-balanced image is converted to a grayscale matrix format and output to the feature recognition module. This step achieves image consistency through color normalization, based on the processing results of the filtered image, and provides standardized data input to the feature recognition module, completing the environmental adaptation processing flow.

[0093] Specifically, the intelligent cleaning system of the present invention, which employs robotic arms for cleaning and drying, further includes a feature recognition module configured as follows:

[0094] Receive the optimized image matrix output by the environment adaptation processing module, perform background segmentation on the optimized image matrix, and distinguish between the part area and the background;

[0095] The segmented image is input into a pre-trained convolutional neural network model, and edge and contour features are extracted through convolutional and pooling layers;

[0096] The extracted features are classified using a fully connected layer to obtain the center coordinates and pose angles of the part; a confidence score is calculated based on the classification data of the center coordinates and pose angles.

[0097] If the confidence score is lower than the preset value, a re-collection command will be triggered;

[0098] If the confidence score is higher than the preset value, the center coordinates, attitude angle, and confidence score are encapsulated into a structured data format and transmitted to the control decision module.

[0099] The feature recognition module receives the optimized image matrix output by the environment adaptation processing module and performs background segmentation on the optimized image matrix to distinguish between part regions and the background. Specifically, the background segmentation uses a threshold segmentation algorithm, setting a dynamic threshold based on the image pixel brightness value to classify pixels into part regions and background regions, generating a binary mask image. This step isolates the target part through image segmentation, providing accurate input for subsequent feature extraction. Logically, it is based on the output of the environment adaptation processing module and prepares processing data for the convolutional neural network model.

[0100] The segmented image is input into a pre-trained convolutional neural network model, which extracts edge and contour features through convolutional and pooling layers. The convolutional layers apply multiple convolutional kernels to scan the image, detecting edge and corner features; the pooling layers perform downsampling operations to reduce the dimensionality of the feature map, preserving salient features and enhancing translation invariance. This process automatically extracts high-level features through the deep learning model, logically relying on the output of background segmentation, and provides feature vectors for classification.

[0101] The extracted features are classified using a fully connected layer to obtain the center coordinates and pose angles of the part. The fully connected layer maps the feature vectors to the output space, calculates the center coordinates of the part through regression, and outputs the pose angles, such as Euler angles or quaternions, through classification. A confidence score is calculated based on the classified data of the center coordinates and pose angles. The confidence score represents the prediction reliability through a softmax function or probability output. This process achieves accurate identification of the part's pose, logically based on feature extraction using a convolutional neural network, and provides a quantitative indicator for decision-making.

[0102] If the confidence score falls below a preset value, a re-acquisition command is triggered. This command is sent to the vision data acquisition module via an interrupt signal or message queue, requesting a rescan of the part to be cleaned. This process avoids low-quality recognition results through a confidence mechanism, logically based on the confidence score evaluation, and forms a feedback loop to improve system robustness.

[0103] If the confidence score is higher than a preset value, the center coordinates, attitude angle, and confidence score are encapsulated into a structured data format and transmitted to the control decision module. The structured data format uses JSON or Protocol Buffers and includes coordinate values, angle values, and confidence values, transmitted via TCP / IP or fieldbus protocols. This step completes the standardized output of the recognition data, logically based on the high-confidence recognition result, and provides input to the control decision module, connecting to subsequent trajectory planning actions.

[0104] Specifically, the intelligent cleaning system of the present invention, which employs robotic arms for cleaning and drying, further includes a control decision module configured as follows:

[0105] Parse the received feature information to obtain the part coordinates and orientation;

[0106] Combining the part coordinates and orientation with the system's preset spray gun incident angle and cleaning time parameters, the inverse kinematics algorithm is used to calculate the motion trajectory of the robotic arm joints;

[0107] After calculating the motion trajectory, the part coordinates and posture are compared with historical recognition data. If the deviation exceeds the preset tolerance value, the coverage of the motion trajectory is adjusted to obtain the adjusted motion trajectory.

[0108] Based on the adjusted motion trajectory, the speed curve of the robotic arm's motion is optimized to generate smooth acceleration commands;

[0109] The motion trajectory and smooth acceleration commands are encapsulated into control signal packets and sent to the robotic arm execution module via the fieldbus protocol.

[0110] The control decision module parses the received feature information to obtain the part's coordinates and orientation. Specifically, the feature information is received from the feature recognition module in a structured data format such as JSON or Protocol Buffers. The parsing operation includes decoding the data packet and extracting the center coordinate values ​​and orientation angle values, such as 3D coordinates and Euler angles. The next step extracts specific pose data from the feature information, providing input for subsequent motion trajectory calculations. Logically, this is based on the output of the feature recognition module and prepares parameters for inverse kinematics calculations.

[0111] Combining the part coordinates and orientation with the system's preset spray gun incident angle and cleaning time parameters, an inverse kinematics algorithm is used to calculate the motion trajectory of the robotic arm joints. The system's preset cleaning parameters include the spray gun incident angle (e.g., the tilt angle relative to the part surface) and the cleaning time (e.g., the spray duration). The inverse kinematics algorithm, based on the desired position and orientation of the robotic arm's end effector, solves for the angle or displacement sequence of each joint, generating the motion path. This step converts the task space coordinates into joint space trajectories through kinematic calculations, logically relying on the parsed part coordinates and orientation, and connecting the trajectory adjustment actions.

[0112] After calculating the motion trajectory, the part coordinates and attitude are compared with historical recognition data. If the deviation exceeds a preset tolerance value, the coverage of the motion trajectory is adjusted to obtain the adjusted motion trajectory. Historical recognition data is stored in the system database, including part coordinate and attitude records from previous recognition cycles. The comparison operation calculates the Euclidean distance between the current data and historical data as the position deviation and the angle difference as the attitude deviation. If the deviation exceeds the tolerance value, such as the position deviation threshold or the angle deviation threshold, the coverage of the motion trajectory is expanded, for example, by adding waypoints or expanding the trajectory bounding box. This process detects and identifies anomalies by comparing historical data, dynamically adjusts the trajectory to compensate for errors, logically based on the inverse kinematics calculation results, and provides input for speed optimization.

[0113] Based on the adjusted motion trajectory, the velocity curve of the robotic arm is optimized to generate smooth acceleration commands. Velocity curve optimization employs a trapezoidal velocity profile or S-curve acceleration algorithm to reduce impact and vibration during robotic arm movement. Smooth acceleration commands are generated through control algorithms such as a proportional-integral-derivative (PID) controller to achieve smooth motion. This step improves motion quality through velocity planning, logically relying on the adjusted motion trajectory, and prepares data for command encapsulation.

[0114] The motion trajectory and smooth acceleration commands are encapsulated into control signal packets and sent to the robotic arm execution module via a fieldbus protocol. The control signal packets are encapsulated using industrial fieldbus protocols such as EtherCAT or PROFINET, and include trajectory data such as position sequences, velocity curves, and acceleration commands, transmitted via a real-time network. This process completes the final generation and transmission of motion commands, logically optimizing the output based on speed, and then connecting to the robotic arm execution module to drive the cleaning action.

[0115] Specifically, the intelligent cleaning system of the present invention employs a robotic arm for cleaning and drying, wherein the robotic arm execution module is further configured as follows:

[0116] Decode the received control signal packets to obtain the target's position, velocity, and acceleration parameters;

[0117] Based on the target position, velocity, and acceleration parameters, drive the servo motor and reducer to control the movement of the robotic arm joints;

[0118] During the movement of the robotic arm joints, the spray gun pressure is adjusted to the preset range of the system, the flow sensor data is monitored, and the joint encoder position data of the robotic arm is collected in real time.

[0119] The joint encoder position data is sent as feedback data to the feature recognition module;

[0120] After receiving the feedback data, the feature recognition module compares the feedback data with the target position of the robotic arm. If the deviation between the feedback data and the target position of the robotic arm exceeds the system's preset tolerance value, it sends a fine-tuning command to the robotic arm execution module.

[0121] The robotic arm execution module receives a fine-tuning instruction, triggering a position fine-tuning mechanism to re-execute the cleaning action.

[0122] The robotic arm execution module decodes the received control signal packets to obtain target position, velocity, and acceleration parameters. The control signal packets are received from the control decision module and encapsulated using fieldbus protocols such as EtherCAT or PROFINET. The decoding operation includes parsing the data packets to extract the target position coordinates, velocity curve values, and acceleration command values. This step completes the parsing of motion commands, providing specific parameters for driving the robotic arm.

[0123] Based on the target position, velocity, and acceleration parameters, servo motors and reducers are driven to control the movement of the robotic arm joints. The servo motors receive velocity command signals, and the reducers amplify the torque to precisely control the rotation angle or linear displacement of each joint of the robotic arm, enabling trajectory tracking by the end effector. This process converts the motion parameters into physical actions, and the cleaning task is executed based on the decoded parameters.

[0124] During the movement of the robotic arm joints, the spray gun pressure is adjusted to the system's preset range, and the flow sensor data is monitored. Pressure regulation is achieved through a proportional valve or electric actuator to maintain stable spray pressure; the flow sensor monitors the cleaning agent flow rate in real time to ensure uniform supply. Simultaneously, the position data of the robotic arm's joint encoders is collected in real time, and the encoders output angle or position signals for each joint. The process synchronously controls the spray gun and monitors the motion status, making real-time adjustments based on the motion execution.

[0125] The joint encoder position data is sent as feedback data to the feature recognition module. The feedback data, including timestamps and position values, is transmitted via communication protocols such as TCP / IP or fieldbus. Upon receiving the feedback data, the feature recognition module compares it with the target position of the robotic arm and calculates the Euclidean distance as the deviation value. If the deviation exceeds a system-preset tolerance value, a fine-tuning command is sent to the robotic arm execution module. This process forms a feedback loop, performing accuracy evaluation based on the collected data.

[0126] The robotic arm's execution module receives a fine-tuning command, triggering a position fine-tuning mechanism to re-execute the cleaning action. This mechanism adjusts joint positions using an incremental motion algorithm, such as making minor corrections to trajectory points or repeating local paths. Re-executing the cleaning action ensures coverage of areas with deviations, improving cleaning integrity. The process corrects motion errors based on feedback, achieving closed-loop control.

[0127] Specifically, the intelligent cleaning system of the present invention, which employs robotic arms for cleaning and drying, further includes:

[0128] The optimized image matrix output by the environment adaptation processing module is used as the input to the convolutional neural network model in the feature recognition module.

[0129] Convolutional neural network models extract features from the optimized image matrix through convolutional layers and pooling layers, and output feature information.

[0130] The optimized image matrix output from the environment adaptation processing module serves as the input to the convolutional neural network model in the feature recognition module. The optimized image matrix, received from the environment adaptation processing module in grayscale matrix format, includes preprocessed image data, such as pixel values ​​after illumination enhancement and noise suppression. The feature recognition module inputs the optimized image matrix into the pre-trained convolutional neural network model, completing the data handover and preparing for feature extraction. The step connects the output of the environment adaptation processing module with the input of the feature recognition module, initiating the feature recognition process based on the results of the preceding image processing.

[0131] Convolutional neural network models extract features from optimized image matrices using convolutional and pooling layers. Convolutional layers use multiple kernels to scan the input image matrix, detecting local features such as edges and textures, and generating feature maps. Pooling layers perform downsampling operations to reduce the dimensionality of the feature maps, preserving salient features and enhancing translation invariance. This process automatically extracts high-level features using a deep learning model, performing hierarchical processing based on the input image matrix to provide feature vectors for classification output.

[0132] The convolutional neural network model outputs feature information, including extracted contour and positional features. This feature information is generated through fully connected layers or the output layer and represented as feature vectors or feature maps, including the geometric properties and spatial relationships of the part. The output feature information is then transmitted to subsequent processing units for classification or regression tasks. This step completes the feature extraction process, based on the processing results of convolutional and pooling layers, and provides data for the final output of the feature recognition module.

[0133] Specifically, the intelligent cleaning system of the present invention, which uses a robotic arm for cleaning and drying, includes a confidence score in the feature information output by the feature recognition module.

[0134] The control decision module receives feature information and executes the following: if the confidence score is higher than the preset value, it generates robotic arm movement instructions based on the feature information.

[0135] If the confidence score is lower than the preset value, a re-acquisition command is triggered or historical recognition data is called to perform error compensation calculation, generating robotic arm movement commands with expanded coverage.

[0136] The feature recognition module outputs feature information including a confidence score, which is calculated after classification by a fully connected layer and represents the reliability index of the feature recognition result. The feature information is encapsulated in a structured data format, including part coordinates, attitude angles, and confidence scores, and is transmitted to the control decision module via a communication protocol. This step completes the quantitative evaluation of the recognition results, providing data support for control decisions.

[0137] After receiving feature information, the control decision module parses the data packet to extract the confidence score. The parsing operation includes decoding structured data, reading the confidence score field, and comparing it with a preset value. This preset value is stored in the system configuration and serves as a threshold for judging recognition quality. This process of receiving and parsing data initiates the decision-making process based on the output of the feature recognition module.

[0138] If the confidence score is higher than a preset value, the control decision module generates robotic arm motion commands based on the feature information. This feature information includes part coordinates and attitude angles, combined with preset cleaning parameters, and uses an inverse kinematics algorithm to calculate the motion trajectory, optimize the velocity curve to generate smooth acceleration commands, and encapsulate them into a control signal packet. This step executes command generation under high confidence conditions, driving the robotic arm based on reliable recognition results.

[0139] If the confidence score falls below a preset value, the control decision module triggers a re-acquisition command or retrieves historical recognition data for error compensation calculation. The re-acquisition command is sent to the vision data acquisition module via an interrupt signal, requesting a rescan of the part. Historical recognition data is retrieved from the system database to collect previously recorded feature information, and data fusion or interpolation calculations are performed to compensate for the current recognition error. After error compensation calculation, a robotic arm motion command with expanded coverage is generated, extending the motion trajectory boundary to address positional uncertainties. This process handles low-confidence recognition results, improving system fault tolerance through re-acquisition or historical data compensation, and ensuring the integrity of the cleaning action.

[0140] Specifically, the intelligent cleaning system of the present invention, which employs robotic arms for cleaning and drying, further includes:

[0141] The robotic arm execution module feeds back the position data collected by the joint encoder to the feature recognition module;

[0142] The feature recognition module receives feedback position data, calculates the Euclidean distance between the feedback position data and the part coordinates obtained in the current recognition cycle as the position deviation, calculates the difference between the feedback position data and the part attitude angle as the attitude deviation, and merges the position deviation and attitude deviation to generate pose deviation data.

[0143] The feature recognition module adjusts the parameters of the convolutional neural network model in the feature recognition module based on the pose deviation data.

[0144] The robotic arm execution module feeds back the position data collected by the joint encoder to the feature recognition module. The joint encoder collects angle or displacement data of each joint of the robotic arm in real time, forming a position data sequence. This position data is encapsulated into data packets via fieldbus protocols such as EtherCAT or PROFINET, including timestamps and joint coordinate values. The feedback data is then transmitted to the feature recognition module, completing the data interaction between the execution and recognition stages. This process enables real-time acquisition and feedback of the robotic arm's actual movement position, providing fundamental data for deviation calculation.

[0145] The feature recognition module receives feedback position data and calculates the Euclidean distance between the feedback position data and the part coordinates obtained in the current recognition cycle as the position deviation. The Euclidean distance calculation is based on the difference in three-dimensional spatial coordinates; the square root of the sum of squares yields a linear deviation value. Simultaneously, the difference between the feedback position data and the part's attitude angle is calculated as the attitude deviation, using either the absolute value of the angle difference or a vector angle. The position deviation and attitude deviation are combined to generate pose deviation data, represented as a vector or matrix, integrating position and attitude error information. This step quantifies the difference between the actual position and the recognized position through geometric calculations, providing input for model adjustment.

[0146] The feature recognition module adjusts the parameters of the convolutional neural network model based on pose deviation data. Parameter adjustment employs the backpropagation algorithm, calculating the gradient of the loss function based on the pose deviation data to update the weights and biases of the convolutional and fully connected layers. The adjustment process is implemented using an optimizer such as stochastic gradient descent to reduce the error between the model output and the true pose. This step utilizes the deviation data to optimize model performance, improve subsequent recognition accuracy, and form an adaptive learning mechanism.

[0147] Secondly, please refer to Figure 1 This invention provides an intelligent cleaning method using a robotic arm for cleaning and drying, applicable to the intelligent cleaning system described above, comprising:

[0148] Step 1: Scan the part to be cleaned to obtain raw image data; perform a self-check to inspect lens cleanliness and light source status; remove oil stains from the lens surface during scanning; dynamically analyze ambient light intensity and adjust exposure parameters when the brightness is below the threshold; transmit raw image data after image acquisition is complete.

[0149] Step 2: Receive the original image data, perform light intensity analysis and stain area detection, calculate the average brightness distribution of the original image data, trigger the contrast enhancement algorithm to adjust the gamma value when the average brightness of the original image data is lower than the threshold, apply adaptive median filtering to identify and replace noisy pixels, and output the optimized image matrix after performing color balance correction.

[0150] Step 3: Receive the optimized image matrix for background segmentation, extract contour features through convolutional and pooling layers, classify and output part coordinates and pose through a fully connected layer, and calculate the confidence score; if the confidence score is lower than the set value, trigger re-acquisition; if it is higher than the set value, encapsulate the coordinates, pose, and confidence score into structured data.

[0151] Step 4: Receive the encapsulated structured data as feature information, parse the part coordinates and attitude, combine the system's preset spray gun incident angle and cleaning time parameters, use the inverse kinematics algorithm to calculate the motion trajectory, compare historical identification data and adjust the trajectory coverage when the deviation exceeds the limit, optimize the speed curve to generate smooth acceleration commands, encapsulate the motion commands and send them.

[0152] Step 5: Decode the control signal packet to obtain motion parameters, drive the movement of the robotic arm joints, adjust the spray gun pressure and monitor the flow data, and collect the joint position data of the robotic arm in real time as feedback data; compare the feedback data with the target position of the robotic arm, and send a fine-tuning command to trigger a re-cleaning action when the deviation between the feedback data and the target position of the robotic arm exceeds the limit.

[0153] Step 6: Based on the feedback data, calculate the Euclidean distance between the feedback data and the part coordinates as the position deviation, calculate the difference between the feedback data and the attitude angle as the attitude deviation, merge to generate pose deviation data, and adjust the parameters of the convolutional neural network model.

[0154] The vision sensor initiates a self-test program to check lens cleanliness and light source status. This program includes capturing a built-in test image, analyzing image contrast and sharpness to assess lens contamination, and simultaneously using a photometer to check if the light source intensity is within its normal operating range. After the self-test, the vision sensor scans the part to be cleaned. During scanning, an automatic cleaning device is periodically activated to remove oil and dirt from the lens surface. This automatic cleaning device uses pneumatic spray cleaning agent. Ambient light intensity is dynamically analyzed during scanning, with a real-time brightness sensor monitoring the brightness. When the brightness falls below a preset threshold, exposure parameters, including exposure time and ISO gain, are automatically adjusted. After image acquisition, raw image data is obtained. The acquired raw image data is transmitted to the environmental adaptation processing module via the industrial Ethernet protocol. The industrial Ethernet protocol encapsulates image data packets for high-speed and reliable data transmission. This process, through self-testing and cleaning, maintains the vision sensor in optimal working condition and adaptively compensates for low-light conditions, providing high-quality raw image data for subsequent processing.

[0155] After receiving the raw image data, the environmental adaptation processing module performs illumination intensity analysis, parses the raw image data, calculates the average brightness of all pixels, and generates a brightness distribution histogram. Based on the average brightness distribution, when the average brightness is below a threshold, a contrast enhancement algorithm is triggered to adjust the image gamma value. The contrast enhancement algorithm improves the visibility of details in dark areas by mapping pixel brightness values ​​through nonlinear transformation. An adaptive median filter is applied to scan image pixels to identify noisy pixels. The adaptive median filter dynamically adjusts the filter window size, replacing noisy pixels with the weighted average of neighboring pixels to suppress dirt interference. Color balance correction is performed on the filtered image. The color balance correction analyzes the color channel distribution deviation of the image and adjusts the gain ratio of the red, green, and blue channels to compensate for color cast. The color-balanced corrected image is converted into a grayscale matrix format and output to the feature recognition module. These steps compensate for environmental interference through image enhancement and filtering, improve image quality, and provide optimized input for feature recognition.

[0156] The feature recognition module receives the optimized image matrix and performs background segmentation. Background segmentation uses a thresholding algorithm, setting a dynamic threshold based on image pixel brightness values ​​to classify pixels into part regions and background regions, generating a binary mask image. The segmented image is input into a pre-trained convolutional neural network model. Multiple convolutional kernels are applied through convolutional layers to scan the image and detect edge and corner features. Pooling layers perform downsampling to reduce feature map dimensionality while preserving significant features. Fully connected layers classify the extracted features and output the part's center coordinates and pose angles (e.g., Euler angles or quaternions). A confidence score is calculated based on the classified data, and the prediction reliability is represented by a softmax function. If the confidence score is lower than a preset value, a re-acquisition command is triggered and sent to the visual data acquisition module via an interrupt signal. If the confidence score is higher than the preset value, the center coordinates, pose angles, and confidence score are encapsulated in a structured data format (e.g., JSON or Protocol Buffers) and transmitted to the control decision module. This process utilizes a deep learning model to extract features and uses a confidence mechanism to filter reliable recognition results, preventing erroneous data from affecting subsequent control.

[0157] The control decision module receives encapsulated structured data as feature information and parses the feature information to obtain part coordinates and attitude data. Combining the part coordinates and attitude with the system's preset spray gun incident angle and cleaning time parameters, an inverse kinematics algorithm is used to calculate the robot arm joint motion trajectory. The inverse kinematics algorithm solves the joint angle sequence based on the desired position and attitude of the robot arm's end effector. After calculating the motion trajectory, the part coordinates and attitude are compared with historical recognition data, which is stored in the system database. The Euclidean distance between the current data and historical data is calculated as the position deviation, and the angle difference is calculated as the attitude deviation. If the deviation exceeds a preset tolerance value, the motion trajectory coverage is adjusted, such as by adding path points or expanding the trajectory bounding box. Based on the adjusted motion trajectory, the robot arm's motion speed curve is optimized. Speed ​​curve optimization uses a trapezoidal velocity profile or S-curve acceleration algorithm to generate smooth acceleration commands. The motion trajectory and smooth acceleration commands are encapsulated into control signal packets and sent to the robot arm execution module via fieldbus protocols such as EtherCAT or PROFINET. This process generates precise commands through motion planning and error compensation, dynamically adjusting the trajectory to address recognition deviations and achieve complete cleaning coverage.

[0158] The robotic arm execution module decodes the received control signal packets to obtain the target position velocity and acceleration parameters. Based on these parameters, it drives servo motors and reducers to control the movement of the robotic arm joints. The servo motors receive velocity command signals and amplify torque through the reducers to achieve precise motion control. During the movement of the robotic arm joints, the spray gun pressure is adjusted to the system's preset range. Pressure regulation is maintained through proportional valves or electric actuators to ensure stable pressure, and monitoring flow sensor data ensures uniform cleaning agent supply. Real-time acquisition of robotic arm joint encoder position data serves as feedback data; the joint encoders output angle or position signals for each joint. This feedback data is sent to the feature recognition module via a communication protocol such as TCP / IP, including timestamps and position values. The feature recognition module receives the feedback data, compares it with the target position of the robotic arm, calculates the Euclidean distance as the deviation value, and sends a fine-tuning command to the robotic arm execution module if the deviation exceeds the system's preset tolerance. The robotic arm execution module receives the fine-tuning command and triggers a position fine-tuning mechanism to re-execute the cleaning action. This mechanism adjusts the joint positions using an incremental motion algorithm. The process executes motion commands and monitors them in real time. By comparing and detecting execution errors through feedback, it corrects them in a timely manner to achieve cleaning accuracy.

[0159] The Euclidean distance between the feedback data and the part coordinates is calculated as the position deviation, and the difference between the feedback data and the attitude angle is calculated as the attitude deviation. The position deviation and attitude deviation are combined to generate pose deviation data. The feature recognition module adjusts the parameters of the convolutional neural network model based on the pose deviation data. Parameter adjustment uses the backpropagation algorithm to calculate the gradient of the loss function based on the pose deviation data, updating the weights and biases of the convolutional and fully connected layers to optimize model performance. This process utilizes feedback data to optimize the recognition model and improves subsequent recognition accuracy through adaptive learning, forming a closed-loop optimization mechanism for the system.

[0160] This invention addresses the technical problem of inaccurate part position detection caused by recognition errors in vision sensors under low light or oily environments through a multi-module collaborative data processing and closed-loop control mechanism. The visual data acquisition module initiates a self-test program to check lens cleanliness and light source status. During scanning, an automatic cleaning device removes oil from the lens surface and dynamically adjusts exposure parameters to acquire high-quality raw image data. The environmental adaptation processing module performs light intensity analysis and stain area detection, driving image enhancement and filtering. A contrast enhancement algorithm adjusts the gamma value to improve the visibility of details in dark areas, adaptive median filtering identifies and replaces noisy pixels to suppress stain interference, and color balance correction compensates for color cast, outputting an optimized image matrix. The feature recognition module uses a convolutional neural network model to extract part contour and position features from the optimized image matrix, outputting coordinate and orientation angle feature information and a confidence score. If the confidence score falls below a set value, a re-acquisition command is triggered to avoid incorrect recognition. If the value exceeds the set value, the data transmission is encapsulated. The control decision module integrates feature information with preset cleaning parameters, uses inverse kinematics algorithm to calculate the motion trajectory, compares historical recognition data, and adjusts the trajectory coverage when the deviation exceeds the limit, generating robotic arm motion commands. The robotic arm execution module drives the spray gun to complete the cleaning action and feeds back the position data collected by the joint encoder to the feature recognition module. It compares the feedback data with the target position and triggers fine-tuning commands to re-execute the cleaning action when the deviation exceeds the limit, forming a closed-loop control to correct the recognition deviation in real time. The system compensates for low lighting conditions through image enhancement, suppresses stain noise through filtering, filters the recognition results through a confidence mechanism, compensates for the motion range with historical data, and corrects execution errors through feedback loops. Ultimately, it achieves accurate detection of part positions and complete cleaning, solving the problems of missed cleaning and uneven cleaning.

[0161] This invention relates to an intelligent cleaning system employing a robotic arm for cleaning and drying. This system addresses the technical problem of inaccurate part position detection caused by recognition errors in vision sensors under low light or oily conditions through multi-module collaborative operation. In intelligent cleaning workshop applications, high-pressure oil pump parts often accumulate thick oil stains or are located in shadowed areas of the equipment. Vision sensors, due to lens contamination or light scattering, cannot accurately identify the edge contours of the parts. The system achieves a complete cleaning process through the following methods:

[0162] The visual data acquisition module initiates a self-test program on the visual sensor to check lens cleanliness and light source status. This program includes capturing a built-in test image, analyzing image contrast and sharpness to assess lens contamination, and simultaneously using a photometer to detect the light source intensity. After the self-test is complete, the module controls the visual sensor to scan the part to be cleaned. During scanning, a periodically activated automatic cleaning device removes oil and dirt from the lens surface using pneumatic spray cleaning agent. Ambient light intensity is dynamically analyzed during scanning, with a real-time brightness sensor monitoring the brightness. When the brightness falls below a threshold, exposure parameters, including exposure time and ISO gain, are automatically adjusted. After image acquisition, raw image data is obtained and transmitted to the environmental adaptation processing module via the industrial Ethernet protocol. The industrial Ethernet protocol encapsulates image data packets for high-speed and reliable data transmission.

[0163] After receiving the raw image data, the environmental adaptation processing module performs illumination intensity analysis, parses the raw image data, calculates the average brightness of all pixels, and generates a brightness distribution histogram. Based on the average brightness distribution, when the average brightness is below a threshold, a contrast enhancement algorithm is triggered to adjust the image gamma value. The contrast enhancement algorithm improves the visibility of details in dark areas by mapping pixel brightness values ​​through nonlinear transformation. An adaptive median filter is applied to scan image pixels to identify noisy pixels. The adaptive median filter dynamically adjusts the filter window size and replaces noisy pixels with the weighted average of neighboring pixels to suppress dirt interference. Color balance correction is performed on the filtered image. The color balance correction analyzes the color channel distribution deviation of the image and adjusts the gain ratio of the red, green, and blue channels to compensate for color cast. The color-balanced corrected image is then converted into a grayscale matrix format and output to the feature recognition module.

[0164] The feature recognition module receives the optimized image matrix and performs background segmentation. Background segmentation uses a thresholding algorithm, setting a dynamic threshold based on image pixel brightness values ​​to classify pixels into part regions and background regions, generating a binary mask image. The segmented image is then input into a pre-trained convolutional neural network model. Multiple convolutional kernels are applied through convolutional layers to scan the image and detect edge and corner features. Pooling layers perform downsampling to reduce feature map dimensionality while preserving significant features. Fully connected layers classify the extracted features and output the part's center coordinates and pose angles (e.g., Euler angles or quaternions). A confidence score is calculated based on the classification data, and the prediction reliability is represented by a softmax function. If the confidence score is lower than a set value, a re-acquisition command is triggered and sent to the visual data acquisition module via an interrupt signal. If the confidence score is higher than the set value, the center coordinates, pose angles, and confidence score are encapsulated in a structured data format (e.g., JSON or Protocol Buffers) and transmitted to the control decision module.

[0165] The control decision module receives encapsulated structured data as feature information and parses this information to obtain part coordinates and attitude data. Combining the part coordinates and attitude with the system's preset spray gun incident angle and cleaning time parameters, an inverse kinematics algorithm is used to calculate the robot arm joint motion trajectory. This algorithm solves for the joint angle sequence based on the desired position and attitude of the robot arm's end effector. After calculating the motion trajectory, the part coordinates and attitude are compared with historical identification data, which is stored in the system database. The Euclidean distance between the current data and historical data is calculated as the position deviation, and the angle difference as the attitude deviation. If the deviation exceeds the tolerance limit, the motion trajectory coverage is adjusted, such as by adding path points or expanding the trajectory bounding box. Based on the adjusted motion trajectory, the robot arm's motion speed curve is optimized. Speed ​​curve optimization uses a trapezoidal velocity profile or S-curve acceleration algorithm to generate smooth acceleration commands. The motion trajectory and smooth acceleration commands are encapsulated into control signal packets and sent to the robot arm execution module via fieldbus protocols such as EtherCAT or PROFINET.

[0166] The robotic arm execution module decodes the received control signal packets to obtain the target position velocity and acceleration parameters. Based on these parameters, it drives servo motors and reducers to control the movement of the robotic arm joints. The servo motors receive velocity command signals and amplify torque through the reducers to achieve precise motion control. During the movement of the robotic arm joints, the spray gun pressure is adjusted to the system's preset range. Pressure regulation is maintained through proportional valves or electric actuators to ensure stable pressure, and monitoring flow sensor data ensures uniform cleaning agent supply. Real-time acquisition of robotic arm joint encoder position data serves as feedback data; the joint encoders output angle or position signals for each joint. This feedback data is sent to the feature recognition module via a communication protocol such as TCP / IP, including timestamps and position values. The feature recognition module receives the feedback data, compares it with the target position of the robotic arm, calculates the Euclidean distance as the deviation value, and sends a fine-tuning command to the robotic arm execution module if the deviation exceeds the system's preset tolerance. The robotic arm execution module receives the fine-tuning command and triggers a position fine-tuning mechanism to re-execute the cleaning action. This mechanism adjusts the joint positions using an incremental motion algorithm.

[0167] The system calculates the Euclidean distance between the feedback data and the part coordinates as the position deviation, and the difference between the feedback data and the attitude angle as the attitude deviation. The position deviation and attitude deviation are then combined to generate pose deviation data. The feature recognition module adjusts the parameters of the convolutional neural network model based on the pose deviation data. Parameter adjustment uses a backpropagation algorithm to calculate the gradient of the loss function based on the pose deviation data, updating the weights and biases of the convolutional and fully connected layers to optimize model performance. The system compensates for low-light conditions through image enhancement, suppresses dirt noise through filtering, filters recognition results using a confidence mechanism, compensates for the motion range using historical data, and corrects execution errors through a feedback loop, forming a closed-loop control to achieve accurate part position detection and complete cleaning.

[0168] Embodiment 1 of the present invention is applied in a smart cleaning workshop scenario;

[0169] High-pressure oil pump parts often accumulate thick oil stains or are located in shaded areas, making it difficult for vision sensors to accurately identify the edges and contours of parts due to lens contamination or light scattering. This embodiment addresses the identification problem in low-light and oily environments by using a visual data acquisition module and an environmental adaptation processing module working together. The visual data acquisition module initiates a self-test program for the vision sensor, checking lens cleanliness and light source status. The self-test program includes capturing a built-in test image, analyzing image contrast and sharpness to assess the degree of lens contamination, and simultaneously using a photometer to check if the light source intensity is within the normal operating range. After the self-test program is complete, the vision sensor is controlled to scan the part to be cleaned. During the scanning process, an automatically cleaning device periodically activated removes oil stains from the lens surface, using a pneumatic spray cleaning agent. The ambient light intensity is dynamically analyzed during scanning, with an ambient light sensor monitoring brightness values ​​in real time. When the brightness falls below a threshold, exposure parameters, including exposure time and ISO gain, are automatically adjusted. After image acquisition, raw image data is obtained. The acquired raw image data is transmitted to the environmental adaptation processing module via the industrial Ethernet protocol, which encapsulates image data packets for high-speed and reliable data transmission. After receiving the raw image data, the environmental adaptation processing module performs illumination intensity analysis, parses the raw image data, calculates the average brightness of all pixels to generate a brightness distribution histogram, and triggers a contrast enhancement algorithm to adjust the image gamma value when the average brightness is below a threshold. The contrast enhancement algorithm improves the visibility of details in dark areas by mapping pixel brightness values ​​through nonlinear transformation. An adaptive median filter is applied to scan image pixels to identify noisy pixels. The adaptive median filter dynamically adjusts the filter window size, replacing noisy pixels with the weighted average of neighboring pixels to suppress dirt interference. Color balance correction is performed on the filtered image. The color balance correction analyzes the color channel distribution deviation and adjusts the gain ratio of the red, green, and blue channels to compensate for color cast. The resulting color-balanced image is converted to a grayscale matrix format and output to the feature recognition module. This embodiment utilizes self-checking and cleaning to maintain the optimal state of the visual sensor, adaptive exposure and image enhancement to compensate for low lighting conditions, and filtering to suppress dirt noise, providing optimized input for subsequent feature recognition and solving the recognition error problem caused by poor original image quality.

[0170] Example 2 is in the intelligent cleaning process;

[0171] Inaccurate part position detection can lead to missed cleaning or uneven cleaning. This embodiment addresses the identification deviation problem by forming a closed-loop control system consisting of a feature recognition module, a control decision module, and a robotic arm execution module. The feature recognition module receives the optimized image matrix output from the environment adaptation processing module and performs background segmentation. Background segmentation uses a threshold segmentation algorithm that sets a dynamic threshold based on the image pixel brightness value, classifying pixels into part regions and background regions to generate a binary mask image. The segmented image is input into a pre-trained convolutional neural network model. Multiple convolutional kernels are applied through convolutional layers to scan the image and detect edge and corner features. Pooling layers perform downsampling operations to reduce the dimension of the feature map while retaining significant features. Fully connected layers classify the extracted features and output the part center coordinates and pose angles, such as Euler angles or quaternions. A confidence score is calculated based on the classification data, and the prediction reliability is represented by a softmax function. If the confidence score is lower than a set value, a re-acquisition command is triggered and sent to the visual data acquisition module via an interrupt signal. If the confidence score is higher than the set value, the center coordinates, pose angles, and confidence score are encapsulated in a structured data format, such as JSON or Protocol Buffers, and transmitted to the control decision module. After receiving feature information, the control decision module parses the part coordinates and attitude data. Combining this with the system's preset spray gun incident angle and cleaning time parameters, it uses an inverse kinematics algorithm to calculate the robot arm joint motion trajectory. The inverse kinematics algorithm solves for the joint angle sequence based on the desired position and attitude of the robot arm's end effector. After calculating the motion trajectory, the part coordinates and attitude are compared with historical recognition data, which is stored in the system database. The Euclidean distance between the current data and historical data is calculated as the position deviation, and the angle difference is calculated as the attitude deviation. If the deviation exceeds the tolerance value, the motion trajectory coverage is adjusted, such as by adding path points or expanding the trajectory bounding box. Based on the adjusted motion trajectory, the robot arm's motion speed curve is optimized. Speed ​​curve optimization uses a trapezoidal velocity profile or S-curve acceleration algorithm to generate smooth acceleration commands. The motion trajectory and smooth acceleration commands are encapsulated into control signal packets and sent to the robot arm execution module via fieldbus protocols such as EtherCAT or PROFINET. The robotic arm execution module decodes control signal packets to obtain target position velocity and acceleration parameters, drives servo motors and reducers to control the movement of the robotic arm joints, adjusts the spray gun spray pressure to the system's preset range during movement, monitors flow sensor data, and collects joint encoder position data in real time as feedback data, which is then sent to the feature recognition module. The feature recognition module compares the feedback data with the target position; if the deviation exceeds the tolerance value, it sends a fine-tuning command, triggering the position fine-tuning mechanism to re-execute the cleaning action. This embodiment uses a confidence mechanism to filter reliable recognition results, trajectory planning and error compensation to generate precise commands, and a feedback closed loop to correct execution deviations in real time, solving the problems of missed cleaning and uneven cleaning caused by inaccurate part position detection.

[0172] The data types involved in this invention include raw image data, optimized image matrices, feature information, robotic arm motion commands, feedback data, and pose deviation data. Raw image data is acquired by scanning the part to be cleaned using a vision sensor, including collected pixel values ​​and brightness information. The optimized image matrix is ​​output by the environment adaptation processing module, including image data after illumination enhancement, noise suppression, and color balance correction, converted into a grayscale matrix format. Feature information is generated by the feature recognition module, including the center coordinates of the part, attitude angles such as Euler angles or quaternions, and a confidence score representing recognition reliability. Robotic arm motion commands are generated by the control decision module, including motion trajectory data, velocity curves, and smooth acceleration commands, encapsulated into control signal packets. Feedback data comes from the robotic arm execution module, including position data collected by the joint encoder, such as joint angles or displacement values, and timestamp information. Pose deviation data is calculated by the feature recognition module, including the Euclidean distance between the feedback data and the part coordinates as the position deviation, and the difference between the feedback data and the attitude angle as the attitude deviation, merging to generate integrated position and attitude error data. Historical identification data is stored in the system database, encompassing part coordinates and orientation records from previous identification cycles, for comparison and compensation calculations. All data is encapsulated in structured formats such as JSON or Protocol Buffers and transmitted using industrial Ethernet or fieldbus protocols, enabling efficient and reliable data exchange and processing between system modules.

[0173] The visual data acquisition module acquires raw image data through a visual sensor and initiates a self-test program to check lens cleanliness and light source status. This program includes capturing a built-in test image, analyzing image contrast and sharpness to assess the degree of lens contamination, and simultaneously using a photometer to check if the light source intensity is within its normal operating range. After the self-test program completes, the module controls the visual sensor to scan the part to be cleaned. During the scanning process, an automatically cleaning device, periodically activated, removes oil and dirt from the lens surface using a pneumatic spray cleaning agent. The module dynamically analyzes ambient light intensity during scanning, monitoring brightness values ​​in real time through an ambient light sensor. When the brightness falls below a preset threshold, the module automatically adjusts exposure parameters, including exposure time and ISO gain, to acquire raw image data. The acquired raw image data is transmitted to the environmental adaptation processing module via an industrial Ethernet protocol. The industrial Ethernet protocol encapsulates image data packets to achieve high-speed and reliable data transmission.

[0174] After receiving the raw image data, the environmental adaptation processing module performs illumination intensity analysis, parses the raw image data, calculates the average brightness of all pixels, and generates a brightness distribution histogram. Based on the average brightness distribution, when the average brightness is below a threshold, a contrast enhancement algorithm is triggered to adjust the image gamma value. The contrast enhancement algorithm improves the visibility of details in dark areas by mapping pixel brightness values ​​through nonlinear transformation. An adaptive median filter is applied to scan image pixels to identify noisy pixels. The adaptive median filter dynamically adjusts the filter window size and replaces noisy pixels with the weighted average of neighboring pixels to suppress dirt interference. Color balance correction is performed on the filtered image. The color balance correction analyzes the color channel distribution deviation of the image and adjusts the gain ratio of the red, green, and blue channels to compensate for color cast. The color-balanced corrected image is then converted into a grayscale matrix format and output to the feature recognition module.

[0175] The feature recognition module receives the optimized image matrix and performs background segmentation. Background segmentation uses a thresholding algorithm, setting a dynamic threshold based on image pixel brightness values ​​to classify pixels into part regions and background regions, generating a binary mask image. The segmented image is then input into a pre-trained convolutional neural network model. Multiple convolutional kernels are applied through convolutional layers to scan the image and detect edge and corner features. Pooling layers perform downsampling to reduce feature map dimensionality while preserving significant features. Fully connected layers classify the extracted features and output the part's center coordinates and pose angles (e.g., Euler angles or quaternions). A confidence score is calculated based on the classification data, and the prediction reliability is represented by a softmax function. If the confidence score is lower than a preset value, a re-acquisition command is triggered and sent to the visual data acquisition module via an interrupt signal. If the confidence score is higher than the preset value, the center coordinates, pose angles, and confidence score are encapsulated in a structured data format (e.g., JSON or Protocol Buffers) and transmitted to the control decision module.

[0176] The control decision module receives encapsulated structured data as feature information and parses the feature information to obtain part coordinates and attitude data. Combining the part coordinates and attitude with the system's preset spray gun incident angle and cleaning time parameters, an inverse kinematics algorithm is used to calculate the motion trajectory of the robotic arm joints. The inverse kinematics algorithm solves the joint angle sequence based on the desired position and attitude of the robotic arm's end effector. After calculating the motion trajectory, the part coordinates and attitude are compared with historical recognition data, which is stored in the system database. The Euclidean distance between the current data and historical data is calculated as the position deviation, and the angle difference is calculated as the attitude deviation. If the deviation exceeds a preset tolerance value, the motion trajectory coverage is adjusted, such as by adding path points or expanding the trajectory bounding box. Based on the adjusted motion trajectory, the robotic arm's motion speed curve is optimized. Speed ​​curve optimization uses a trapezoidal velocity profile or S-curve acceleration algorithm to generate smooth acceleration commands. The motion trajectory and smooth acceleration commands are encapsulated into control signal packets and sent to the robotic arm execution module via fieldbus protocols such as EtherCAT or PROFINET.

[0177] The robotic arm execution module decodes the received control signal packets to obtain the target position velocity and acceleration parameters. Based on these parameters, it drives servo motors and reducers to control the movement of the robotic arm joints. The servo motors receive velocity command signals and amplify torque through the reducers to achieve precise motion control. During the movement of the robotic arm joints, the spray gun pressure is adjusted to the system's preset range. Pressure regulation is maintained through proportional valves or electric actuators to ensure stable pressure, and monitoring flow sensor data ensures uniform cleaning agent supply. Real-time acquisition of robotic arm joint encoder position data serves as feedback data; the joint encoders output angle or position signals for each joint. This feedback data is sent to the feature recognition module via a communication protocol such as TCP / IP, including timestamps and position values. The feature recognition module receives the feedback data, compares it with the target position of the robotic arm, calculates the Euclidean distance as the deviation value, and sends a fine-tuning command to the robotic arm execution module if the deviation exceeds the system's preset tolerance. The robotic arm execution module receives the fine-tuning command and triggers a position fine-tuning mechanism to re-execute the cleaning action. This mechanism adjusts the joint positions using an incremental motion algorithm.

[0178] The system calculates the Euclidean distance between the feedback data and the part coordinates as the position deviation, and the difference between the feedback data and the attitude angle as the attitude deviation. The position deviation and attitude deviation are then combined to generate pose deviation data. The feature recognition module adjusts the parameters of the convolutional neural network model based on the pose deviation data. Parameter adjustment uses a backpropagation algorithm to calculate the gradient of the loss function based on the pose deviation data, updating the weights and biases of the convolutional and fully connected layers to optimize model performance. The system compensates for low-light conditions through image enhancement, suppresses dirt noise through filtering, filters recognition results using a confidence mechanism, compensates for the range of motion using historical data, and corrects execution errors through a feedback loop, forming a closed-loop control to achieve accurate part position detection and complete cleaning.

[0179] Convolutional Neural Networks (CNNs) are deep learning architectures used to process image data. They automatically extract local features such as edges and textures through convolutional layers, pooling layers reduce feature dimensionality and enhance translation invariance, and fully connected layers integrate features for classification or regression output. In this invention, the CNN model is constructed as a pre-trained model, comprising multiple stacked convolutional and pooling layers to progressively abstract image features, and finally connected to a fully connected layer for final decision-making. The model is trained on a large amount of labeled data to optimize weight parameters and learn the mapping relationship between part contours and positional features. The CNN model receives an optimized image matrix from the environment adaptation processing module as input data. This image matrix undergoes illumination enhancement, noise suppression, and color balance correction, and is converted to grayscale format to focus shape information. The model's data processing begins with the image region after background segmentation. The input image is first scanned by multiple convolutional kernels through convolutional layers to detect edge and corner features and generate a feature map. Subsequently, pooling layers perform downsampling operations to compress the data, retaining key features and reducing computational complexity. The extracted feature vectors are passed to fully connected layers, which map the features to the output space through nonlinear transformations. This performs a classification task to identify the part category and regresses to calculate precise coordinate values ​​and pose angles. The convolutional neural network model output includes the part's center coordinates (e.g., 3D spatial position), pose angles (e.g., Euler angles or quaternions representing direction), and a confidence score calculated based on the output to indicate recognition reliability. The output data is encapsulated as structured feature information and transmitted to the control decision module for subsequent motion trajectory planning and cleanup control. This processing flow achieves end-to-end feature learning and pose estimation, improving the system's recognition accuracy and robustness in complex environments.

[0180] The contrast enhancement algorithm is applied to the raw image data processing in the environment adaptation processing module. When the average brightness is lower than a preset threshold after parsing the raw image data and calculating the average brightness distribution, the algorithm is triggered. The algorithm adjusts the image gamma value through nonlinear transformation and maps the pixel brightness value to enhance the visibility of details in dark areas and compensate for the contrast loss caused by low lighting conditions. The processed image data improves the saliency of local features, provides optimized input for subsequent filtering operations, and connects with the adaptive median filtering algorithm to further suppress noise.

[0181] The adaptive median filtering algorithm processes the contrast-enhanced image data in the environmental adaptation processing module. The algorithm dynamically adjusts the size of the filtering window to scan image pixels, identifies noisy pixels such as stains or random interference points, and replaces them with the weighted average of neighboring pixels. The process effectively smooths the image texture, preserves edge information, and suppresses stain noise, outputting filtered image data. The filtering result prepares a clean input for the color balance correction algorithm, ensuring the accuracy of subsequent color processing.

[0182] The color balance correction algorithm processes the image output by adaptive median filtering in the environmental adaptation processing module, analyzes the distribution deviation of the red, green and blue channels of the image, and adjusts the gain ratio of each channel to compensate for the color cast caused by oil stains. After correction, the image has a uniform color distribution, restores the true color tone, and is finally converted into a grayscale matrix format for output to the feature recognition module. This step completes the image preprocessing process and provides a standardized data foundation for feature extraction.

[0183] The inverse kinematics algorithm is used in the control decision module to calculate the motion trajectory. It receives the part coordinates and attitude angle feature information output by the feature recognition module, and combines them with the system's preset spray gun incident angle and cleaning time parameters. The algorithm solves the joint angle or displacement sequence according to the expected position and attitude of the robotic arm end effector to generate a precise motion path. The calculated motion trajectory is used for subsequent error compensation comparison and is connected to historical data deviation analysis to dynamically adjust the coverage range.

[0184] The confidence score calculation algorithm processes the part center coordinates and attitude angle data output by the fully connected layer classification in the feature recognition module, and calculates the predicted reliability score through the softmax function or probability output model. The confidence score quantifies the credibility of the recognition result. If the score is lower than the set threshold, a re-acquisition instruction is triggered to avoid erroneous recognition. If the score is higher than the threshold, the data is encapsulated and transmitted to the control decision module. This mechanism improves the robustness of the system and ensures that only high-quality data drives subsequent decisions.

[0185] The Euclidean distance algorithm is used for deviation calculation in multiple modules of the system. In the control decision module, the spatial distance between the current part coordinates and historical recognition data is compared, and the square root of the sum of squares is used to obtain the linear position deviation value. In the feature recognition module, the Euclidean distance between the feedback joint encoder position data and the target position is calculated as the basis for evaluating the execution error. The deviation result is used to determine whether the limit is exceeded and to trigger trajectory adjustment or fine-tuning commands to correct the motion error.

[0186] In the feature recognition module, the backpropagation algorithm adjusts the parameters of the convolutional neural network model based on the pose deviation data. The pose deviation data is generated by merging the Euclidean distance and pose angle difference between the feedback position data and the target coordinates. The algorithm calculates the gradient of the loss function and updates the weights and biases of the convolutional and fully connected layers through an optimizer such as stochastic gradient descent to optimize the model's recognition accuracy. The adaptive learning process continuously improves the accuracy of feature extraction, forming a closed-loop optimization mechanism to enhance system performance.

Claims

1. An intelligent cleaning system employing robotic arms for cleaning and drying, characterized in that, include; The visual data acquisition module is configured to scan the parts to be cleaned using a visual sensor installed at the end of the robotic arm, thereby acquiring raw image data. The environmental adaptation processing module is connected to the visual data acquisition module and is configured to receive raw image data, perform light intensity analysis and stain area detection on the raw image data to obtain analysis results and stain area detection results, use the analysis results and stain area detection results to drive image enhancement and filtering processing, and output an optimized image matrix. The feature recognition module is connected to the environment adaptation processing module and is configured to receive the optimized image matrix, extract the contour and position features of the part from the optimized image matrix, and output feature information including coordinates and attitude angles. The control decision module is connected to the feature recognition module and is configured to receive feature information, call the system's preset cleaning parameters based on the feature information, fuse the feature information with the preset cleaning parameters, perform trajectory planning based on the fused data to calculate the robotic arm's motion trajectory, perform error compensation calculation to correct motion errors, and generate robotic arm motion commands. The robotic arm execution module is connected to the control decision module and is configured to execute motion commands to drive the spray gun to complete the cleaning action. During the driving process, the joint encoder collects position data in real time and feeds the position data back to the feature recognition module. The robotic arm execution module is also configured as follows: Decode the received control signal packets to obtain the target's position, velocity, and acceleration parameters; Based on the target position, velocity, and acceleration parameters, drive the servo motor and reducer to control the movement of the robotic arm joints; During the movement of the robotic arm joints, the spray gun pressure is adjusted to the preset range of the system, the flow sensor data is monitored, and the joint encoder position data of the robotic arm is collected in real time. The joint encoder position data is sent as feedback data to the feature recognition module; After receiving the feedback data, the feature recognition module compares the feedback data with the target position of the robotic arm. If the deviation between the feedback data and the target position of the robotic arm exceeds the system's preset tolerance value, it sends a fine-tuning command to the robotic arm execution module. The robotic arm execution module receives a fine-tuning command, triggering a position fine-tuning mechanism to re-execute the cleaning action; The feature recognition module is also configured as follows: Receive the optimized image matrix output by the environment adaptation processing module, perform background segmentation on the optimized image matrix, and distinguish between the part area and the background; The segmented image is input into a pre-trained convolutional neural network model, and edge and contour features are extracted through convolutional and pooling layers; The extracted features are classified using a fully connected layer to obtain the center coordinates and attitude angles of the part; a confidence score is calculated based on the classification data of the center coordinates and attitude angles. If the confidence score is lower than the preset value, a re-collection command will be triggered; If the confidence score is higher than the preset value, the center coordinates, attitude angle and confidence score are encapsulated into a structured data format and transmitted to the control decision module; Also includes: The robotic arm execution module feeds back the position data collected by the joint encoder to the feature recognition module; The feature recognition module receives feedback position data, calculates the Euclidean distance between the feedback position data and the part coordinates obtained in the current recognition cycle as the position deviation, calculates the difference between the feedback position data and the part attitude angle as the attitude deviation, and merges the position deviation and attitude deviation to generate pose deviation data. The feature recognition module adjusts the parameters of the convolutional neural network model in the feature recognition module based on the pose deviation data.

2. The intelligent cleaning system employing robotic arms for cleaning and drying according to claim 1, characterized in that, The visual data acquisition module is also configured as follows: Initiate the self-test program of the vision sensor to check the lens cleanliness and light source status; After the self-test is completed, the vision sensor is controlled to scan the parts to be cleaned. During the scanning process, the oil and dirt on the lens surface are removed by the periodically activated automatic cleaning device. During the scanning process, the ambient light intensity is dynamically analyzed, and the exposure parameters are automatically adjusted when the brightness is below the threshold to complete image acquisition; The acquired raw image data is transmitted to the environmental adaptation processing module via the industrial Ethernet protocol.

3. The intelligent cleaning system employing robotic arms for cleaning and drying according to claim 2, characterized in that, The environmental adaptation processing module is also configured as follows: The received raw image data is parsed, and the average brightness distribution of the raw image data is calculated. Based on the average brightness distribution, when the average brightness of the original image data is lower than the threshold, the contrast enhancement algorithm is triggered to adjust the gamma value of the original image data. An adaptive median filter is applied to scan the pixels of the original image data, identify noisy pixels in the image and replace them with the mean of neighboring pixels to obtain the filtered image. The filtered image is subjected to color balance correction to obtain a color balance corrected image. The color balance corrected image is then converted into a grayscale matrix format and output to the feature recognition module.

4. The intelligent cleaning system employing robotic arms for cleaning and drying according to claim 3, characterized in that, The control decision module is also configured as follows: Parse the received feature information to obtain the part coordinates and orientation; Combining the part coordinates and orientation with the system's preset spray gun incident angle and cleaning time parameters, the inverse kinematics algorithm is used to calculate the motion trajectory of the robotic arm joints; After calculating the motion trajectory, the part coordinates and posture are compared with historical recognition data. If the deviation exceeds the preset tolerance value, the coverage of the motion trajectory is adjusted to obtain the adjusted motion trajectory. Based on the adjusted motion trajectory, the speed curve of the robotic arm's motion is optimized to generate smooth acceleration commands; The motion trajectory and smooth acceleration commands are encapsulated into control signal packets and sent to the robotic arm execution module via the fieldbus protocol.

5. The intelligent cleaning system employing robotic arms for cleaning and drying according to claim 4, characterized in that, Also includes: The optimized image matrix output by the environment adaptation processing module is used as the input to the convolutional neural network model in the feature recognition module. Convolutional neural network models extract features from the optimized image matrix through convolutional layers and pooling layers, and output feature information.

6. The intelligent cleaning system employing robotic arm cleaning and drying according to claim 5, characterized in that, The feature information output by the feature recognition module includes a confidence score; The control decision module receives feature information and executes the following: if the confidence score is higher than the preset value, it generates robotic arm movement instructions based on the feature information. If the confidence score is lower than the preset value, a re-acquisition command is triggered or historical recognition data is called to perform error compensation calculation, generating robotic arm movement commands with expanded coverage.

7. A smart cleaning method employing robotic arms for cleaning and drying, applied to the smart cleaning system employing robotic arms for cleaning and drying as described in any one of claims 1 to 6, characterized in that, include: Step 1: Scan the part to be cleaned to obtain raw image data; perform a self-check to inspect lens cleanliness and light source status; During the scanning process, oil and dirt are removed from the lens surface; the ambient light intensity is dynamically analyzed and the exposure parameters are adjusted when the brightness is below the threshold; and the raw image data is transmitted after image acquisition is completed. Step 2: Receive the original image data, perform light intensity analysis and stain area detection, calculate the average brightness distribution of the original image data, trigger the contrast enhancement algorithm to adjust the gamma value when the average brightness of the original image data is lower than the threshold, apply adaptive median filtering to identify and replace noisy pixels, and output the optimized image matrix after performing color balance correction. Step 3: Receive the optimized image matrix for background segmentation, extract contour features through convolutional and pooling layers, classify and output part coordinates and pose through a fully connected layer, and calculate the confidence score; if the confidence score is lower than the set value, trigger re-acquisition; if it is higher than the set value, encapsulate the coordinates, pose, and confidence score into structured data. Step 4: Receive the encapsulated structured data as feature information, parse the part coordinates and attitude, combine the system's preset spray gun incident angle and cleaning time parameters, use the inverse kinematics algorithm to calculate the motion trajectory, compare historical identification data and adjust the trajectory coverage when the deviation exceeds the limit, optimize the speed curve to generate smooth acceleration commands, encapsulate the motion commands and send them. Step 5: Decode the control signal packet to obtain motion parameters, drive the movement of the robotic arm joints, adjust the spray gun pressure and monitor the flow data, and collect the joint position data of the robotic arm in real time as feedback data. The system compares the feedback data with the target position of the robotic arm. When the deviation between the feedback data and the target position of the robotic arm exceeds the limit, a fine-tuning command is sent to trigger a re-cleaning action. Step 6: Based on the feedback data, calculate the Euclidean distance between the feedback data and the part coordinates as the position deviation, calculate the difference between the feedback data and the attitude angle as the attitude deviation, merge to generate pose deviation data, and adjust the parameters of the convolutional neural network model.

Images